CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[chalsall]

Originally Posted by Batalov

I could use one...

Just be aware they run "warm" (read: consume a lot of energy and require a lot of space).

Jerry gave me a 580 a few years ago, for which I will be forever grateful.

[chalsall]

Originally Posted by flashjh

I have a couple of 580s sitting around that are very stable for all work. I can send them to anyone who would like them.

Might I suggest that you send those 580s to those who use electricity for heating during the winter months?

Compute is free if, and only if, the heat is needed rather than being expended.

[airsquirrels]

Originally Posted by chalsall

Might I suggest that you send those 580s to those who use electricity for heating during the winter months?

Compute is free if, and only if, the heat is needed rather than being expended.

Replacing the heating bill with the computing bill is a favorite of mine. The moment when we pulled out the resistive heating furnace from one of our buildings and replaced it with a liquid-cooling-loop radiator from several GPU boxes for the Ohio winter was quite satisfying.

[Batalov]

Originally Posted by chalsall

Just be aware they run "warm" (read: consume a lot of energy and require a lot of space)..

I have a couple 570s-580s, so yes, I know. One of them is in the process of slowly dying , so it would be a replacement for it...

(I've repaired it a few times, but it already hear its maker's call... and of course I am not trading it via warranty - because EVGA warranty dept usually without asking you 1. receives the card, 2. finds that it's not passing tests, and 3. sends you some lowball replacement like a 750 not even a Ti )

[kladner]

Originally Posted by chalsall

Just be aware they run "warm" (read: consume a lot of energy and require a lot of space).

Jerry gave me a 580 a few years ago, for which I will be forever grateful.

244 W at stock. Mine is not at stock. I'm not sure just how much the 580 is using.

Just checked the one measure I have: Kil-a-Watt. With CPU and GPUs all running, the 580 doing mfaktc, it reads ~830 W. Stop the 580, it drops to ~545W. I can't test the idling watts, without pulling the card. However, clearly mfaktc (plus harder running fans) is sucking down 285 W. ATM, it is running at 918 MHZ, with the memory turned down to 1400 to save power and reduce heat.

Turning it down to 900 MHz, dropped K-a-W to ~810 W, after the fans slowed down. The core temp dropped from 77 to 74 C.

At my hot weather setting of 861 MHz, it reaches ~775 W, 72-73 C.

Again, these are whole system readings, so only the changes can be attributed to the GPU.[/CAPTAIN OBVIOUS]

[kladner]

Originally Posted by airsquirrels

Replacing the heating bill with the computing bill is a favorite of mine. The moment when we pulled out the resistive heating furnace from one of our buildings and replaced it with a liquid-cooling-loop radiator from several GPU boxes for the Ohio winter was quite satisfying.

That is a triumph of conservation!
Here is a way to get cooling from that heat:
https://en.wikipedia.org/wiki/Absorption_refrigerator
I have known these as "gas refrigerators." The gas, in this case, is the household natural gas used to apply heat to the evaporator regenerator.

[storm5510]

Originally Posted by kladner

...Here is a question: have you run CUFFTbench and threadbench on this card?

No, I have not. I did not know about the option until I did some further reading. I will give it a try.

[storm5510]

A screen capture is below of the threadbench test. Perhaps this may shed some light on this issue.

[TheJudger]

Linux, CUDALucas 2.05.1, CUDA 8.0 running './CUDALucas -cufftbench 2592 8192 20'

Device              Tesla P100-PCIE-16GB
Compatibility       6.0
clockRate (MHz)     1328
memClockRate (MHz)  715

  fft    max exp  ms/iter
 2592   48471289   0.7706
 2744   51250889   0.8155
 3136   58404433   0.9263
 3200   59570449   1.1191
 3240   60298969   1.1376
 3888   72075517   1.1380
 4000   74106457   1.1456
 4096   75846319   1.1500
 4116   76208701   1.4355
 5184   95507747   1.4920
 5488  100984691   1.5897
 5600  103000823   1.9282
 5760  105879517   1.9856
 5832  107174381   2.0210
 6000  110194363   2.0535
 6048  111056879   2.0923
 6075  111541967   2.1055
 6144  112781477   2.1056
 6272  115080019   2.1215
 6400  117377567   2.1458
 6480  118813021   2.1929
 6561  120266023   2.2281
 7776  142017539   2.2600
 8192  149447533   2.2731

And not even close the TDP limit: Testing M74207281 with 4000K FFT yields ~175W power draw (card only) while TDP is 250W.

Oliver

[ATH]

Wow, that is ~32 hours for a 80M exponent and ~75 hours for 120M very nice!

And ~22 hours for a 70M compared to ~62 hours on a Titan Black and ~52 hours for 8 cores on a Haswell-E 5960X with quad channel DDR4.

What is the price on a Tesla P100 PCI-E ?



Edit: And ~24 hours on M74207281 compared to Madpoo's "record" of 34 hours on a dual Xeon v3 with 20 cores on it.

[ATH]

Now there is a Nvidia Quadro GP100 coming in March:
http://www.anandtech.com/show/11102/...s-quadro-gp100

Unfortunately they only state FP64 = 1/2 FP32, not the actual numbers.

The long and short of it is that the Quadro GP100 is meant to be a Tesla/GP100 card for workstations, but with even more functionality. While NVIDIA offers PCIe Tesla P100 cards, those cards only feature passive cooling and are designed for servers; the lack of active cooling means you can’t put them in (conventional) workstations. The Quadro GP100 on the other hand is a traditional, fan & shroud active cooled card, like the rest of the Quadro lineup. And then NVIDIA doesn’t stop there, enabling graphics functionality that isn’t on the Tesla cards.

		    Quadro GP100	  Tesla P100 PCI-E
CUDA Cores		3584
Stream Processors				3584
Texture Units		224
ROPs			128?
Core Clock		 ?			  ?
Boost Clock	       1430MHz			1300Mhz
Memory Clock	    1.4 Gbps HBM2	    1.4 Gbps HBM2
Memory Bus Width       4096 bit		       4096 bit
Memory Bandwidth1			       720 GB/sec
VRAM			16GB			16GB
ECC			Full			Full
FP16			 ?		      18.7 TFLOPS
FP32			 ?		       9.3 TFLOPS
FP64			1/2 FP32	       4.7 TFLOPS
GPU			GP100			GP100
TDP			235W			250W
Transistor Count				15.3B
Architecture		Pascal			Pascal